Summary: This paper examines the use of random matrix theory as it has been applied to model large financial datasets, especially for the purpose of estimating the bias inherent in mean-variance portfolio allocation when a sample covariance matrix is substituted for the true underlying covariance. Such problems were observed and modeled in the seminal work of L. Laloux et al. [“Noise dressing of financial correlation matrices”, Phys. Rev. Lett. 83, No. 7, 1467–1470 (1999; doi:10.1103/physrevlett.83.1467)] and rigorously proved by Z. Bai et al. [Math. Finance 19, No. 4, 639–667 (2009; Zbl 1185.91155)] under minimal assumptions. If the returns on assets to be held in the portfolio are assumed independent and stationary, then these results are universal in that they do not depend on the precise distribution of returns. This universality has been somewhat misrepresented in the literature, however, as asymptotic results require that an arbitrarily long time horizon be available before such predictions necessarily become accurate. In order to reconcile these models with the highly non-Gaussian returns observed in real financial data, a new ensemble of random rectangular matrices is introduced, modeled on the observations of independent Lévy processes over a fixed time horizon.